Research on Fusion Algorithm of Hyper Spectral and High Spatial Resolution Remote Sensing Image
|School||Harbin Institute of Technology|
|Course||Information and Communication Engineering|
|Keywords||Hyperspectral image images fusion feature extraction injection model|
With the rapid development of remote sensing technology and the continuous advancement of new sensors, the capability of obtaining data on remote sensing picture increases evidently. We can get a lot of information on different scales, different spectrum and different time in the same area. Hyperspectral data and high spatial resolution image are the most widely used ones of all the data. It’s too hard to improve both spatial and spectral resolution at the same time for one sensor, and the spatial and spectral information in the later application are very significant. So how to utilize the two types of data synthetically to extract more complete accurate information is a key subject in the remote sensing image processing. Remote sensing image fusion is one of the most effective way to solve the problem, it will be two complementary data, combined and have a more complete and effective data, applications for the latter part of a more accurate and reliable data. It can apperceive two types of data’s advantages and disadvantages and make the data complementary each other and we get a group of more integrated and effective data, so in later application, more accurate and credible data is applied. First, we research on the feature extraction algorithms of hyper spectral data. Then we use the extracted image data and high spatial resolution images to study the pixel image fusion classic algorithms, and the improved method, and basis of this, an injection model for pixel-level fusion method to achieve the unification of the pixel-level fusion.First of all, in-depth analysis of hyperspectral data features and later applications, it is demonstrated feature extraction is highly needed. The linear feature extraction and the study of nonlinear feature extraction algorithms based on the manifold on highspectral data are studied, and we focus on the PCA transform based on the linear feature extraction and LLE study of nonlinear feature extraction algorithms. We research on the principle and realization process, and we use the classical RX detection algorithm and classification algorithm based on Bayesian maximum likelihood to process the result after feature extraction to verify the prepotency of the methods, then the conclusion is that the linear feature extraction methods is suitbale for the detection and the other is suitbale for classification.Then, based on the data after feature extraction on hyperspectral data and high spatial resolution image, the research of second part is the pixel-level fusion of remote sensing image. The principle of several classic pixel fusion algorithm including IHS, PCA, SCN, Brovey and HPF are studied first. Based on this, there is some spectrum distortion in the result of IHS, SCN fusion results. So we propose the improvement of these two methods. After the analysis of theoretical based on improved methods, showing the concrete steps, the simulation is carried out through. By comparing the results of statistical parameters, the improved method of the IHS, SCN algorithm show more satisfactory results on spectrum and spatial imfromation.Finally, by comparing the principles and formulas of those classic fusion algorithms on pixel-level image fusion and evolving the formulas of them, the uniform formula is acquired. There are some parameters in the uniform formula. The different classical method corresponds to the combination of different parameters. So we can find the best pixel-level fusion result by get the best combination of parameters. Then we proposed a fusion algorithm based on injection model. This algorithm is based on the parameters of choice, scale transformation, transmission parameters of changing scales, and genetic algorithm. The new algorithm has a very good performanse through the result of experiments.